How do you transcribe Japanese, Chinese or Korean audio?

You can transcribe all three languages with Whisper, an open model you can run on your own computer, or with a built-in option for some of them, as the table below shows. The steps match those for other languages: set the spoken language, transcribe the whole recording, and check the output against the audio. Two things differ. Chinese and Japanese have no spaces between words, so the usual score for transcripts does not fit them well. And Chinese has two writing systems, so you must pick the right one. For the basics, see our overview of transcription, and for the other languages in this guide, see transcribing and translating audio in other languages.

All three are in Whisper's language list, and Google Cloud supports Japanese

Whisper's language table includes Chinese, Korean and Japanese, each with its own code ¹. Google Cloud's Speech-to-Text supported-languages table lists Japanese (Japan) for its chirp_2 model ². Google's service is an API for developers, not an app you use directly.

Being in a list is not the same as being accurate. For Japanese, the Whisper paper reports a figure of 5.3 percent on the FLEURS read-speech test for its large-v2 model ³. Read that figure with the next section in mind. We did not verify a Chinese or Korean figure from the same test, so this page does not quote one.

Word error rate does not fit Chinese and Japanese, so those scores measure characters

Word error rate counts wrong words, and Chinese and Japanese do not mark word boundaries with spaces. The Whisper paper handled this by putting a space between every letter for Chinese, Japanese, Thai, Lao and Burmese. The authors say this effectively measures the character error rate instead ⁴. So the 5.3 percent Japanese figure is a character-level score on read speech, not a count of wrong words.

The boundaries themselves are hard to agree on. In tests with native Mandarin speakers, their agreement on where the word boundaries fall was about 75 percent ⁵, which is why researchers usually use character error rate for Chinese. For the term, see character error rate. If a tool reports a word error rate for Chinese or Japanese, ask what it counted as a word.

Where the main tools stand for these languages

The table below shows what the sources above support. Blank cells mean we did not check that combination.

Apple's Dictation list for the Mac includes Mandarin Chinese for mainland China and Mandarin for Taiwan ⁷. Microsoft's list of languages that support voice typing in Windows 11 includes Japanese and Korean ⁸.

Choose the Chinese variant before you transcribe

Chinese has more than one variant, and the choice changes the output. Apple's auto-punctuation list for Mac Dictation names Chinese (Simplified), Chinese (Traditional) and Cantonese (Taiwan) as separate entries ⁹. Choose the setting that matches the output you need, and read the first minute to confirm the characters. Speaker region is not always a reliable guide to the script you will get.

Steps to transcribe a recording

1. Set the language: Japanese, Korean, or Chinese with the right variant (Mandarin or Cantonese, Simplified or Traditional).
2. Transcribe the first minute or two and read it against the audio. Check that the script is the one you chose and that the words make sense in context.
3. Check names, numbers and dates against the audio. In Japanese and Chinese, a single wrong character can change a word, so read these closely.
4. Read the whole transcript rather than scoring it by word count. The word boundary problem above means a word count tells you little.
5. Check any English words and brand names separately. Mixed-language recordings are usually harder for speech models ¹⁰.

If you need the text in English, translate the checked transcript, not the raw output. The question on translating a voice recording to English gives the steps.

What to do with the result

For a recording that matters, such as an interview you will quote, have a fluent speaker check the passages that carry the meaning. For live dictation on a computer, use the built-in options in the table, and test them with your own speech before you rely on them.

For the other languages in this guide, see how to transcribe Hindi audio and whether AI can transcribe African languages. For the score used in most transcript checks, see word error rate.